Hyperspectral images clustering on reconfigurable hardware using the k-means algorithm

Unsupervised clustering is a powerful technique for understanding multispectral and hyperspectral images, k-means being one of the most used iterative approaches. It is a simple though computationally expensive algorithm, particularly for clustering large hyperspectral images into many categories. Software implementation presents advantages such as flexibility and low cost for implementation of complex functions. However, it presents limitations, such as difficulties in exploiting parallelism for high performance applications. In order to accelerate the k-means clustering, a hardware implementation could be used. The disadvantage in this approach is that any change in the project requires previous knowledge of the hardware design process and can take several weeks to be implemented. In order to improve the design methodology, an automatic and parameterized implementation for hyperspectral images has been developed in a hardware/software codesign approach. An unsupervised clustering technique k-means that uses the Euclidian distance to calculate the pixel to centers distance was used as a case study to validate the methodology. Two implementations, a software and a hardware/software codesign one, have been implemented. Although the hardware component operates at 40 MHz, being 12.5 times less than the software operating frequency (PC), the codesign implementation was approximately 2 times faster than software one.